Repository navigation
Recreate Qwen3.5 B300 AgentX recipes for a full sweep / 重建 Qwen3.5 B300 AgentX 配方并运行完整测试 - #3931
functionstackx wants to merge 3 commits into
Conversation
Recreate PR #2978 at 7614284 from origin/main, preserving its recipe and synthetic acceptance behavior. Append its performance changelog entry after the current history for a fresh full sweep. 中文：基于最新远端 main 重新创建 PR #2978，保留原有 Qwen3.5 B300 AgentX 配方与合成接受行为，并在当前性能日志末尾追加对应条目，以便重新运行完整扫描。 Co-authored-by: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com> Co-authored-by: Ankur-singh <ankusingh@nvidia.com>
中文：将复制的 AgentX 完整扫描条目关联到新建的 PR #3931。
|
Thanks for the contribution!
中文感谢你的贡献！
|
functionstackx
added
the
full-sweep-enabled
|
Generally most InferenceX submissions don't need to add changes to the infx package. On the off chance they do, generally the diff is small. Please prompt your agent to clean your PR to keep your diff to only necessary & minimal changes too. We have an increased code quality bar for changes to the infx package. 中文大多数 InferenceX 提交通常不需要修改 infx 包。即使确实需要，改动通常也很小。请让你的 agent 清理 PR，使改动只包含必要且最小的内容。infx 包的改动有更高的代码质量要求。 |
Remove SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE from all four B300 AgentX recipes and update this PR performance changelog entry. 中文：从全部四个 B300 AgentX 配方中删除 SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE，并更新本 PR 的性能变更日志。
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=38289020106 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38289020106 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| SGLANG_FLASHINFER_FP4_GEMM_BACKEND: cutlass | ||
| SGLANG_MOE_NVFP4_DISPATCH: '1' | ||
| SGLANG_CUTEDSL_MOE_NVFP4_DISPATCH: '1' | ||
| SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE: '1' |
There was a problem hiding this comment.
🔴 These new B300 AgentX recipes enable SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE=1, which CONTRIBUTING.md:71-79 explicitly and unconditionally prohibits (precision hacking of the MTP draft head), so merging ships a policy-violating config that a reviewer following the checklist (inferencex-e2e/docs/PR_REVIEW_CHECKLIST.md:44) must reject. The PR description itself admits the conflict is 'unresolved and must be addressed before merge,' but the code still sets the flag in all four new recipes. Fix: unset/disable SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE (or get an explicit, recorded policy exception) in every recipe that sets it before merge.
Why this was flagged
CONTRIBUTING.md:71-79 states enabling SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is 'explicitly prohibited going forward, including inherited environment, launcher, container, or image defaults' and 'historical runs are not precedent or an exception.' All four new recipe files set this to '1': agg-tp2-b2-mtp5-hicache.yaml:47, agg-tp2-mtp7-hicache-variants.yaml:48, agg-tp2ep2-mtp3-hicache.yaml:47, agg-tp4-mtp6-hicache-variants.yaml:48. On base main these recipes did not exist; merging introduces four configs that violate the current rule, lowering effective MTP draft precision below the shipped checkpoint default while results are reported as compliant AgentX submissions. The PR_REVIEW_CHECKLIST.md:44 'Explicitly verified... not enabled' gate is the safeguard, but it only catches this if a human reviewer manually checks each file; nothing in the diff itself enforces or blocks the setting.
Verification: CONTRIBUTING.md:71-79 explicitly prohibits enabling SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE, and reviewers must verify it is disabled; historical runs are not precedent. All four newly added recipes set the prohibited value '1': agg-tp2-b2-mtp5-hicache.yaml:47, agg-tp2-mtp7-hicache-variants.yaml:48, agg-tp2ep2-mtp3-hicache.yaml:47, agg-tp4-mtp6-hicache-variants.yaml:48. On base main these four files did not exist, so merging introduces four policy-violating configs where none existed.
This branch has not been deployed
1 participant
Summary
Recreate #2978 (source commit
761428402083ded38ae4908e33f7bd6c9b0ef3ee) on the latest fetchedmain(20c8fba9e) for an independent full sweep. Carry over its five Qwen3.5 NVFP4 B300 Dynamo-SGLang AgentX configurations, pinnedlmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729image, and synthetic-acceptance validation.The matching shared SquashFS cache on B300 was removed so the image can be imported afresh. Use
full-sweep-enabledfor the new run, including the normal canary and selected evals.Difference from #2978: remove
SGLANG_NVFP4_CKPT_FP8_NEXTN_MOEfrom all four recipes, as requested, instead of carrying over the prohibited enabling override. The remaining recipe settings and synthetic-acceptance behavior are unchanged.AI model disclosure
GPT-6 prepared this duplicate, performed the cache cleanup, and ran local verification. The runtime does not expose a verifiable exact model/version identifier, so the exact version could not be verified. No delegated agents contributed. The source PR does not provide verifiable identities for any models used to prepare its original changes.
PR checklist and change type
Type of Change
Checklist
inferencex-e2e/perf-changelog.yamland have not edited historical entriesfull-sweep-fail-fast,full-sweep-enabled, ornon-canary-full-sweep-enabled. Optional modifiersall-evals,evals-only, andagentx-fastrequire a primary label; the last two block reuse while applied.OWNER/MEMBER/COLLABORATOR) has commented/use <run_id>(or the legacy/reuse-sweep-run) on this PR. Only do this after a final full sweep is all green with evals passing.中文
基于最新获取的
main（20c8fba9e）重新创建 #2978（源提交761428402083ded38ae4908e33f7bd6c9b0ef3ee），以独立运行一次完整 sweep。沿用原有五组 Qwen3.5 NVFP4 B300 Dynamo-SGLang AgentX 配置、固定镜像lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729及合成接受校验逻辑。改动涵盖配方配置、校验逻辑和文档。已删除 B300 上对应的共享 SquashFS 缓存，以便重新导入镜像。新运行使用
full-sweep-enabled标签，保留正常的 canary 和默认选定的评测。与 #2978 的区别： 按要求从全部四个配方中删除
SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE，不再沿用被规范禁止的启用设置。其余配方配置和合成接受行为保持不变。AI 模型披露：由 GPT-6 完成 PR 复制、缓存清理及本地验证。运行环境未提供可核实的精确模型或版本标识，因此无法确认精确版本。未使用委派代理。源 PR 未提供可核实的原始改动所用模型标识。